kernel/rcutiny.c

   1 /*
   2  * Read-Copy Update mechanism for mutual exclusion, the Bloatwatch edition.
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  17  *
  18  * Copyright IBM Corporation, 2008
  19  *
  20  * Author: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
  21  *
  22  * For detailed explanation of Read-Copy Update mechanism see -
  23  *              Documentation/RCU
  24  */
  25 #include <linux/completion.h>
  26 #include <linux/interrupt.h>
  27 #include <linux/notifier.h>
  28 #include <linux/rcupdate.h>
  29 #include <linux/kernel.h>
  30 #include <linux/export.h>
  31 #include <linux/mutex.h>
  32 #include <linux/sched.h>
  33 #include <linux/types.h>
  34 #include <linux/init.h>
  35 #include <linux/time.h>
  36 #include <linux/cpu.h>
  37 #include <linux/prefetch.h>
  38
  39 #ifdef CONFIG_RCU_TRACE
  40 #include <trace/events/rcu.h>
  41 #endif /* #else #ifdef CONFIG_RCU_TRACE */
  42
  43 #include "rcu.h"
  44
  45 /* Forward declarations for rcutiny_plugin.h. */
  46 struct rcu_ctrlblk;
  47 static void invoke_rcu_callbacks(void);
  48 static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp);
  49 static void rcu_process_callbacks(struct softirq_action *unused);
  50 static void __call_rcu(struct rcu_head *head,
  51                        void (*func)(struct rcu_head *rcu),
  52                        struct rcu_ctrlblk *rcp);
  53
  54 #include "rcutiny_plugin.h"
  55
  56 static long long rcu_dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
  57
  58 /* Common code for rcu_idle_enter() and rcu_irq_exit(), see kernel/rcutree.c. */
  59 static void rcu_idle_enter_common(long long oldval)
  60 {
  61         if (rcu_dynticks_nesting) {
  62                 RCU_TRACE(trace_rcu_dyntick("--=",
  63                                             oldval, rcu_dynticks_nesting));
  64                 return;
  65         }
  66         RCU_TRACE(trace_rcu_dyntick("Start", oldval, rcu_dynticks_nesting));
  67         if (!is_idle_task(current)) {
  68                 struct task_struct *idle = idle_task(smp_processor_id());
  69
  70                 RCU_TRACE(trace_rcu_dyntick("Error on entry: not idle task",
  71                                             oldval, rcu_dynticks_nesting));
  72                 ftrace_dump(DUMP_ALL);
  73                 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
  74                           current->pid, current->comm,
  75                           idle->pid, idle->comm); /* must be idle task! */
  76         }
  77         rcu_sched_qs(0); /* implies rcu_bh_qsctr_inc(0) */
  78 }
  79
  80 /*
  81  * Enter idle, which is an extended quiescent state if we have fully
  82  * entered that mode (i.e., if the new value of dynticks_nesting is zero).
  83  */
  84 void rcu_idle_enter(void)
  85 {
  86         unsigned long flags;
  87         long long oldval;
  88
  89         local_irq_save(flags);
  90         oldval = rcu_dynticks_nesting;
  91         WARN_ON_ONCE((rcu_dynticks_nesting & DYNTICK_TASK_NEST_MASK) == 0);
  92         if ((rcu_dynticks_nesting & DYNTICK_TASK_NEST_MASK) ==
  93             DYNTICK_TASK_NEST_VALUE)
  94                 rcu_dynticks_nesting = 0;
  95         else
  96                 rcu_dynticks_nesting  -= DYNTICK_TASK_NEST_VALUE;
  97         rcu_idle_enter_common(oldval);
  98         local_irq_restore(flags);
  99 }
 100 EXPORT_SYMBOL_GPL(rcu_idle_enter);
 101
 102 /*
 103  * Exit an interrupt handler towards idle.
 104  */
 105 void rcu_irq_exit(void)
 106 {
 107         unsigned long flags;
 108         long long oldval;
 109
 110         local_irq_save(flags);
 111         oldval = rcu_dynticks_nesting;
 112         rcu_dynticks_nesting--;
 113         WARN_ON_ONCE(rcu_dynticks_nesting < 0);
 114         rcu_idle_enter_common(oldval);
 115         local_irq_restore(flags);
 116 }
 117
 118 /* Common code for rcu_idle_exit() and rcu_irq_enter(), see kernel/rcutree.c. */
 119 static void rcu_idle_exit_common(long long oldval)
 120 {
 121         if (oldval) {
 122                 RCU_TRACE(trace_rcu_dyntick("++=",
 123                                             oldval, rcu_dynticks_nesting));
 124                 return;
 125         }
 126         RCU_TRACE(trace_rcu_dyntick("End", oldval, rcu_dynticks_nesting));
 127         if (!is_idle_task(current)) {
 128                 struct task_struct *idle = idle_task(smp_processor_id());
 129
 130                 RCU_TRACE(trace_rcu_dyntick("Error on exit: not idle task",
 131                           oldval, rcu_dynticks_nesting));
 132                 ftrace_dump(DUMP_ALL);
 133                 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 134                           current->pid, current->comm,
 135                           idle->pid, idle->comm); /* must be idle task! */
 136         }
 137 }
 138
 139 /*
 140  * Exit idle, so that we are no longer in an extended quiescent state.
 141  */
 142 void rcu_idle_exit(void)
 143 {
 144         unsigned long flags;
 145         long long oldval;
 146
 147         local_irq_save(flags);
 148         oldval = rcu_dynticks_nesting;
 149         WARN_ON_ONCE(rcu_dynticks_nesting < 0);
 150         if (rcu_dynticks_nesting & DYNTICK_TASK_NEST_MASK)
 151                 rcu_dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
 152         else
 153                 rcu_dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
 154         rcu_idle_exit_common(oldval);
 155         local_irq_restore(flags);
 156 }
 157 EXPORT_SYMBOL_GPL(rcu_idle_exit);
 158
 159 /*
 160  * Enter an interrupt handler, moving away from idle.
 161  */
 162 void rcu_irq_enter(void)
 163 {
 164         unsigned long flags;
 165         long long oldval;
 166
 167         local_irq_save(flags);
 168         oldval = rcu_dynticks_nesting;
 169         rcu_dynticks_nesting++;
 170         WARN_ON_ONCE(rcu_dynticks_nesting == 0);
 171         rcu_idle_exit_common(oldval);
 172         local_irq_restore(flags);
 173 }
 174
 175 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 176
 177 /*
 178  * Test whether RCU thinks that the current CPU is idle.
 179  */
 180 int rcu_is_cpu_idle(void)
 181 {
 182         return !rcu_dynticks_nesting;
 183 }
 184 EXPORT_SYMBOL(rcu_is_cpu_idle);
 185
 186 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
 187
 188 /*
 189  * Test whether the current CPU was interrupted from idle.  Nested
 190  * interrupts don't count, we must be running at the first interrupt
 191  * level.
 192  */
 193 int rcu_is_cpu_rrupt_from_idle(void)
 194 {
 195         return rcu_dynticks_nesting <= 0;
 196 }
 197
 198 /*
 199  * Helper function for rcu_sched_qs() and rcu_bh_qs().
 200  * Also irqs are disabled to avoid confusion due to interrupt handlers
 201  * invoking call_rcu().
 202  */
 203 static int rcu_qsctr_help(struct rcu_ctrlblk *rcp)
 204 {
 205         if (rcp->rcucblist != NULL &&
 206             rcp->donetail != rcp->curtail) {
 207                 rcp->donetail = rcp->curtail;
 208                 return 1;
 209         }
 210
 211         return 0;
 212 }
 213
 214 /*
 215  * Record an rcu quiescent state.  And an rcu_bh quiescent state while we
 216  * are at it, given that any rcu quiescent state is also an rcu_bh
 217  * quiescent state.  Use "+" instead of "||" to defeat short circuiting.
 218  */
 219 void rcu_sched_qs(int cpu)
 220 {
 221         unsigned long flags;
 222
 223         local_irq_save(flags);
 224         if (rcu_qsctr_help(&rcu_sched_ctrlblk) +
 225             rcu_qsctr_help(&rcu_bh_ctrlblk))
 226                 invoke_rcu_callbacks();
 227         local_irq_restore(flags);
 228 }
 229
 230 /*
 231  * Record an rcu_bh quiescent state.
 232  */
 233 void rcu_bh_qs(int cpu)
 234 {
 235         unsigned long flags;
 236
 237         local_irq_save(flags);
 238         if (rcu_qsctr_help(&rcu_bh_ctrlblk))
 239                 invoke_rcu_callbacks();
 240         local_irq_restore(flags);
 241 }
 242
 243 /*
 244  * Check to see if the scheduling-clock interrupt came from an extended
 245  * quiescent state, and, if so, tell RCU about it.  This function must
 246  * be called from hardirq context.  It is normally called from the
 247  * scheduling-clock interrupt.
 248  */
 249 void rcu_check_callbacks(int cpu, int user)
 250 {
 251         if (user || rcu_is_cpu_rrupt_from_idle())
 252                 rcu_sched_qs(cpu);
 253         else if (!in_softirq())
 254                 rcu_bh_qs(cpu);
 255         rcu_preempt_check_callbacks();
 256 }
 257
 258 /*
 259  * Invoke the RCU callbacks on the specified rcu_ctrlkblk structure
 260  * whose grace period has elapsed.
 261  */
 262 static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp)
 263 {
 264         char *rn = NULL;
 265         struct rcu_head *next, *list;
 266         unsigned long flags;
 267         RCU_TRACE(int cb_count = 0);
 268
 269         /* If no RCU callbacks ready to invoke, just return. */
 270         if (&rcp->rcucblist == rcp->donetail) {
 271                 RCU_TRACE(trace_rcu_batch_start(rcp->name, 0, 0, -1));
 272                 RCU_TRACE(trace_rcu_batch_end(rcp->name, 0,
 273                                               ACCESS_ONCE(rcp->rcucblist),
 274                                               need_resched(),
 275                                               is_idle_task(current),
 276                                               rcu_is_callbacks_kthread()));
 277                 return;
 278         }
 279
 280         /* Move the ready-to-invoke callbacks to a local list. */
 281         local_irq_save(flags);
 282         RCU_TRACE(trace_rcu_batch_start(rcp->name, 0, rcp->qlen, -1));
 283         list = rcp->rcucblist;
 284         rcp->rcucblist = *rcp->donetail;
 285         *rcp->donetail = NULL;
 286         if (rcp->curtail == rcp->donetail)
 287                 rcp->curtail = &rcp->rcucblist;
 288         rcu_preempt_remove_callbacks(rcp);
 289         rcp->donetail = &rcp->rcucblist;
 290         local_irq_restore(flags);
 291
 292         /* Invoke the callbacks on the local list. */
 293         RCU_TRACE(rn = rcp->name);
 294         while (list) {
 295                 next = list->next;
 296                 prefetch(next);
 297                 debug_rcu_head_unqueue(list);
 298                 local_bh_disable();
 299                 __rcu_reclaim(rn, list);
 300                 local_bh_enable();
 301                 list = next;
 302                 RCU_TRACE(cb_count++);
 303         }
 304         RCU_TRACE(rcu_trace_sub_qlen(rcp, cb_count));
 305         RCU_TRACE(trace_rcu_batch_end(rcp->name, cb_count, 0, need_resched(),
 306                                       is_idle_task(current),
 307                                       rcu_is_callbacks_kthread()));
 308 }
 309
 310 static void rcu_process_callbacks(struct softirq_action *unused)
 311 {
 312         __rcu_process_callbacks(&rcu_sched_ctrlblk);
 313         __rcu_process_callbacks(&rcu_bh_ctrlblk);
 314         rcu_preempt_process_callbacks();
 315 }
 316
 317 /*
 318  * Wait for a grace period to elapse.  But it is illegal to invoke
 319  * synchronize_sched() from within an RCU read-side critical section.
 320  * Therefore, any legal call to synchronize_sched() is a quiescent
 321  * state, and so on a UP system, synchronize_sched() need do nothing.
 322  * Ditto for synchronize_rcu_bh().  (But Lai Jiangshan points out the
 323  * benefits of doing might_sleep() to reduce latency.)
 324  *
 325  * Cool, huh?  (Due to Josh Triplett.)
 326  *
 327  * But we want to make this a static inline later.  The cond_resched()
 328  * currently makes this problematic.
 329  */
 330 void synchronize_sched(void)
 331 {
 332         rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
 333                            !lock_is_held(&rcu_lock_map) &&
 334                            !lock_is_held(&rcu_sched_lock_map),
 335                            "Illegal synchronize_sched() in RCU read-side critical section");
 336         cond_resched();
 337 }
 338 EXPORT_SYMBOL_GPL(synchronize_sched);
 339
 340 /*
 341  * Helper function for call_rcu() and call_rcu_bh().
 342  */
 343 static void __call_rcu(struct rcu_head *head,
 344                        void (*func)(struct rcu_head *rcu),
 345                        struct rcu_ctrlblk *rcp)
 346 {
 347         unsigned long flags;
 348
 349         debug_rcu_head_queue(head);
 350         head->func = func;
 351         head->next = NULL;
 352
 353         local_irq_save(flags);
 354         *rcp->curtail = head;
 355         rcp->curtail = &head->next;
 356         RCU_TRACE(rcp->qlen++);
 357         local_irq_restore(flags);
 358 }
 359
 360 /*
 361  * Post an RCU callback to be invoked after the end of an RCU-sched grace
 362  * period.  But since we have but one CPU, that would be after any
 363  * quiescent state.
 364  */
 365 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
 366 {
 367         __call_rcu(head, func, &rcu_sched_ctrlblk);
 368 }
 369 EXPORT_SYMBOL_GPL(call_rcu_sched);
 370
 371 /*
 372  * Post an RCU bottom-half callback to be invoked after any subsequent
 373  * quiescent state.
 374  */
 375 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
 376 {
 377         __call_rcu(head, func, &rcu_bh_ctrlblk);
 378 }
 379 EXPORT_SYMBOL_GPL(call_rcu_bh);